Status as of 19 November 2026: Qwen 4 is partly out. Qwen4-Max launched on 16 November as a hosted, closed-weight API model (2.4T parameters, 71B active, 1M context, $1.50 in / $4.50 out per 1M tokens). Open-weight sizes: none yet; Qwen says smaller open models are "coming to open source", with no sizes, licence or date. Qwen 4 27B: not announced. Ollama: ollama run qwen4:27b available? No (ollama.com/library/qwen4 returns 404). Run qwen3.8:27b today.
| On 19 November 2026 | Last changed | |
|---|---|---|
| Qwen4-Max (hosted) | Released 16 Nov 2026: API via QwenCloud, Alibaba Cloud Model Studio, OpenRouter and others | 16 Nov |
| Open-weight Qwen 4 sizes | None. Qwen: "smaller models… coming to open source". No names, licence or date | 16 Nov |
| Qwen 4 27B | Not announced by Qwen. Hugging Face "Qwen4-27B" repos are third-party placeholders | — |
Ollama qwen4 tag | Not available | — |
| Official benchmarks | Qwen4-Max vs Qwen3.8-Max: yes, below. Qwen 4 27B: none | 16 Nov |
We check Qwen's Hugging Face organisation, the Ollama library, Qwen's blog and the QwenLM GitHub every weekday and change this block the same day anything moves. Changelog at the bottom.
Picture this: you search Hugging Face for Qwen4-27B and get a full page of hits, all uploaded the same evening.
Finally.
You open one. Two files: a .gitattributes and a README that says placeholder, no weights yet.
Is Qwen 4 out? Max yes, open weights no
Two different questions now. If you want to call Qwen 4 over an API, yes: Qwen4-Max has been live since 16 November. If you want to download Qwen 4 and run it on your GPU, no: Qwen has published no Qwen 4 weights, and the only sentence it has written about them is that smaller models are "coming to open source, built on the same lineage and trained with the same generation of techniques".
You can check it yourself in ten seconds:
# Qwen's own Hugging Face org: [] means no Qwen4 repos yet
curl -s "https://huggingface.co/api/models?author=Qwen&search=Qwen4"
# Ollama library page: 404 means it is not there yet
curl -s -o /dev/null -w "%{http_code}\n" https://ollama.com/library/qwen4
# Third: OpenRouter already lists the hosted model (returns the model object)
curl -s https://openrouter.ai/api/v1/models | grep -o '"id":"qwen/qwen4-max"'
Look at the publisher, not just the name. Anyone can create a repo called Qwen4-27B, and the only Qwen 4 that counts comes from the Qwen account.
What Qwen has confirmed about Qwen 4
Alibaba's press release from the Apsara Conference, dated 22 September 2026, said Qwen 4 "is currently in training." It named no sizes, no date and no licence. On 16 November Qwen published Qwen4-Max in Qwen4-Max: Your New Teammate Has Arrived, with the specs below, and wrote that smaller open-weight models are "coming to open source".
What Qwen and Alibaba Cloud's model page say about Qwen4-Max:
- Size: 2.4T total parameters, 71B active per token (Qwen3.8-Max had 95B active), plus 328B n-gram embedding parameters.
- Architecture: Gated DeltaNet layers with Qwen Sparse Attention, gated residual branches and the n-gram embedding: the design Flash Next previewed.
- Context: 1M tokens. Output up to 131,072 tokens without thinking and 262,144 in thinking mode, where
max_tokenscovers thinking and answer together. - Input: text, image and video.
- Defaults: thinking on,
reasoning_efforthigh (low, high and xhigh are accepted),preserve_thinkingon. - Weights: closed. It is an API model only.
The September release also sketched what comes after: the Qwen 4.5 and Qwen 5 series are projected to scale to 5 to 10 trillion parameters. For scale, the biggest open Qwen model today, Qwen3.8-2.4T-A95B, has 2.4 trillion.
The other concrete Qwen 4 fact sits on a model card. Qwen's Hugging Face card for Qwen3.8-Flash-Next calls it an "experimental preview of the architecture that will underpin Qwen4." It runs 125B parameters with only 6B active per token, and Qwen4-Max now confirms that design made it into the flagship.
If you want to know how an open Qwen 4 will behave, don't wait for the launch post. Flash Next is the preview, and you can download it today.
Our Gemma 4 vs Qwen 3.8 breakdown covers Flash Next's benchmarks and memory needs next to every Gemma 4 size.
What about open weights? Qwen has now put it in writing for the smaller sizes: they "are coming to open source". Alibaba's press release says "open-weight, smaller-scale models from the Qwen4 series will soon be released." The flagship is closed. Rumoured dates for the open sizes are everywhere; none of them come from Qwen.
What Qwen's own code suggested, and got right
Before launch, Qwen Code, Qwen's open-source coding agent, merged "fix(core): size the qwen4 generation at 1M" (PR #15762, 28 October 2026). Its tests named qwen4, qwen4-max, qwen4-plus, qwen4-coder-plus and qwen4.5-max. One of those ids was right: qwen4-max shipped with a 1M window, and on 17 November Qwen Code merged full support for it, with its effort ladder, cache TTL and image input (PR #17513). The others, qwen4-plus and qwen4-coder-plus, are still unshipped, and Qwen Code's own design notes treat other qwen4 routes as having no vendor evidence yet. Read them as code references, not a product list.
Qwen 4 sizes and the 27B question
One official size: Qwen4-Max (2.4T, 71B active, hosted). Open sizes: none named as of 19 November 2026. Press reports list Qwen4-Plus, Qwen4-Flash and Qwen4-27B, but Qwen's published materials name none of them. If you're choosing a 27B-class model right now, our Gemma 4 31B vs Qwen 3.8 27B comparison runs the two dense models side by side.
The 27B guess has history behind it, though. Here's what Qwen has shipped as open weights this year:
| Generation | Open sizes on Qwen's Hugging Face | Released |
|---|---|---|
| Qwen 3.5 | 0.8B, 2B, 4B, 9B, 27B, 35B-A3B, 122B-A10B, 397B-A17B | February 2026 |
| Qwen 3.6 | 27B, 35B-A3B | April 2026 |
| Qwen 3.7 | None (hosted Qwen3.7-Plus only) | Hosted only |
| Qwen 3.8 | 27B, Flash Next (125B, 6B active), 2.4T-A95B | August 2026 |
| Qwen 4 | None yet. Qwen4-Max is hosted and closed; open sizes "coming to open source" | 16 Nov 2026 (Max only) |
27B is the only size in every open generation this year. It's also the size that fits a 24GB card: Ollama's default qwen3.8:27b download is 18GB. That's why people search for a Qwen 4 27B, and why we'd bet on one too. A bet, not a fact: as of 19 November 2026, Qwen has not named a Qwen 4 27B.

One warning about those Qwen4-27B uploads on Hugging Face. They come from a third-party account, they're tagged placeholder, and there are no weights inside. Don't build a download script around them.
Until Qwen names sizes, plan from what each tier of hardware runs today and what it would need if Qwen 4 follows the 3.8 pattern:
| Your GPU / memory | Runs today | If Qwen 4 ships a dense 27B (assumption, not announcement) | More |
|---|---|---|---|
| 8–12 GB | qwen3.5:9b, gemma4:12b, Bonsai 2 27B (ternary, 8.3GB) | A 4-bit 27B would not fit (about 17–18GB); wait for a 9B-class or a ternary build | Best coding model for 12GB VRAM |
| 16 GB | A 27B spills to CPU; the 9B–14B class fits | Same | Best local LLM for 16GB VRAM |
| 24 GB | qwen3.8:27b 18GB, gemma4:31b 19–20GB | The target tier for a 27B at q4_K_M | Gemma 4 31B vs Qwen 3.8 27B |
| 32 GB | qwen3.8:27b-q8_0 30GB | q8 27B | — |
| 64 GB+ (Mac or workstation) | qwen3.8:27b-bf16 56GB | bf16 27B | — |
| 128 GB-class | qwen3.8-flash-next 105GB to 120GB | A Flash-Next-sized Qwen 4 MoE, if one ships | Gemma 4 vs Qwen 3.8 |
For the smaller sizes across both families, our Gemma 4 vs Qwen 3.5 comparison is the hub.
Qwen 4 benchmarks vs Qwen 3.8
One like-for-like comparison exists today: Qwen4-Max against Qwen3.8-Max-0902, both hosted, both from Alibaba's own launch chart. There is no Qwen 4 27B, so the open-weight column is empty and stays empty until Qwen publishes a model card. Alibaba's numbers are vendor-reported on Alibaba's harness; the Artificial Analysis and Vals rows are independent and use the same yardstick across models. Read direction, not decimals.
| Benchmark (what it measures) | Qwen4-Max | Qwen3.8-Max-0902 | Qwen 3.8 27B (open) | Qwen 4 27B (open) | Measured by |
|---|---|---|---|---|---|
| Artificial Analysis Intelligence Index v5 | 43 | 34 | 20 | — | Artificial Analysis, independent |
| AA cost per Index task | $3.27 | $10.02 | $3.53 | — | Artificial Analysis |
| Vals Index | 63.20% (#14/42) | 59.05% | — | — | Vals AI, independent |
| Terminal-Bench 4.0 (agentic terminal coding) | 48.5 | 38.9 | — | — | Alibaba chart; Vals measured 48.48% |
| DeepSWE v1.1 (agentic coding) | 75.7 | 69.3 | — | — | Alibaba chart |
| CodeArena Elo (web development) | 1750 | 1676 | — | — | Arena, cited by Alibaba |
| Agents' Last Exam v1.1 (long-horizon tasks) | 46.7 | 27.9 | — | — | ALE leaderboard, cited by Alibaba |
| AutomationBench-AA (agentic SaaS workflow) | 72.5 | 56.2 | — | — | Artificial Analysis, cited by Alibaba |
| APEX-Agents v1.1 (professional services) | 72.1 | 63.3 | — | — | Mercor, cited by Alibaba |
| JobBench (professional job tasks) | 68.5 | 65.3 | — | — | Alibaba chart (updated dataset) |
| AA-Briefcase v1.1 Elo (agentic knowledge work) | 1686 | 1605 | — | — | Artificial Analysis, cited by Alibaba |
| HLAB-Vals-1.1 (legal agent work) | 21.3 | 15.4 | — | — | Vals (21.25% measured) |
| Humanity's Last Exam | 51.7 | 43.9 | — | — | Alibaba chart |
| HealthBench Professional | 60.0 | 39.6 | — | — | Alibaba chart |
| Chartography (visual chart understanding) | 54.2 | 42.0 | — | — | Alibaba chart |
| Work360 (Alibaba's own, 721 fields) | 52.69 | 47.74 | — | — | Alibaba, in-house |
| Cost per Work360 task | $1.67 | $4.43 | — | — | Alibaba, in-house |
| Active parameters | 71B | 95B | 27B (dense) | — | Qwen |
| Price per 1M in / out | $1.50 / $4.50 | $2.00 / $6.00 | $0.50 / $3.00 (hosted) or free locally | — | Alibaba Cloud; AA |
Checked 19 November 2026.
Three things to notice. First, the biggest jumps are on long-horizon agent work (Agents' Last Exam 27.9 → 46.7, AutomationBench 56.2 → 72.5, HealthBench 39.6 → 60.0), which is what the Qwen4 architecture was built for. Second, on the two coding benchmarks everyone quotes, Qwen4-Max leads its own chart on DeepSWE but sits well behind Kimi K3.1 (67.2) and Opus 5.5 (59.6) on Terminal-Bench 4.0. Third, the cost story is real but has a catch: Artificial Analysis measured $3.27 per task against $10.02 for Qwen3.8-Max, and also 146 seconds to first token and the most verbose output in its class. Cheaper per task, slower per turn.
Do not read across to the 27B. The Qwen 3.8 27B model card reports Terminal Bench 2.1 (73.0) and SWE-bench Pro (61.7), different benchmark versions from the Max chart. When a Qwen 4 open model ships, we'll fill the column from its card, and only on rows that use the same version.
Qwen 4 vs Qwen 3.8: what would make it worth upgrading
Qwen 3.8 already set a high bar. On Qwen's own model cards, the 27B went from 63.4 to 73.0 on Terminal Bench 2.1 and from 53.5 to 61.7 on SWE-bench Pro between Qwen 3.6 and 3.8. These are vendor-reported numbers on Qwen's own setup, so read them as direction, not decimals.
Benchmarks aren't the whole story, though. Thinking is the quiet cost: Qwen 3.8 keeps reasoning from every earlier turn by default, and in a 30-step agent loop that piles up fast. If your agent stalls or balloons on Qwen 3.8 today, read our Qwen 3.8 27B tool calling fix first, because most of those problems will follow you to Qwen 4.
So what would Qwen 4 need to be worth the switch for an agent? Here's our checklist, with what Qwen4-Max now tells us:
| Area | Qwen 3.8 today | Worth upgrading if Qwen 4... | What Qwen4-Max shows |
|---|---|---|---|
| Context | 262K native, extensible to about 1M | Gives you 1M natively on the open sizes | 1M, on the hosted model |
| Long-context speed | 27B is a dense model | Brings the Flash Next attention to the 27B class | Flash Next attention is in Max |
| Tool calling | 27B: 67.1 on Toolathlon Verified | Matches Flash Next's 73.5 at 27B size | No public Toolathlon number yet |
| Thinking | On by default, preserve_thinking on | Spends fewer tokens per agent step | max_tokens now covers thinking and answer, so budget 32K or more |
| Licence | 27B Apache 2.0, Flash Next custom | Keeps Apache 2.0 on the sizes you run | Max is closed; open sizes unknown |
The Flash Next numbers are the tell. With 6B active parameters it scores 62.5 on SWE-bench Pro and 73.5 on Toolathlon Verified, against 61.7 and 67.1 for the dense 27B. The architecture Qwen 4 is built on already does more with less.
The question for Qwen 4 isn't "is it smarter?" It's "does it fit my GPU, keep the licence I need, and call tools without eating my context?"
| Qwen 3.8 27B | Qwen3.8-Flash-Next | Qwen3.8-Max-0902 | Qwen4-Max | Qwen 4 open sizes | |
|---|---|---|---|---|---|
| Weights | Open | Open | Hosted (open sibling: 2.4T-A95B) | Closed | Promised, unnamed |
| Licence | Apache 2.0 | Qwen Community 1.0 | — | — | Unknown |
| Context | 262K native, YaRN to 1M | 256K | 1M | 1M | Unknown |
| Active params | 27B | 6B of 125B | 95B of 2.4T | 71B of 2.4T | Unknown |
| Input | Text, image, video | Text, image, video | Text, image | Text, image, video | Unknown |
Checked 19 November 2026.
What to run today: Qwen 3.8 27B, 3.8 Flash Next or Gemma 4 31B
Waiting costs you weeks of agent work for open weights with no date. Here's what Ollama serves right now, plus the one Qwen 4 you can call:
| Model | Ollama tag | Download | Context | Licence |
|---|---|---|---|---|
| Qwen 3.8 27B | qwen3.8:27b | 18GB | 256K | Apache 2.0 |
| Qwen 3.8 Flash Next | qwen3.8-flash-next:125b-a6b-q4_K_M | 120GB | 256K | Qwen Community |
| Gemma 4 31B | gemma4:31b | 19GB to 20GB | 256K | Apache 2.0 |
| Qwen4-Max | API only, qwen4-max | $1.50 / $4.50 per 1M | 1M | Closed |
Qwen4-Max is the only Qwen 4 you can use today, and only over an API: see Try Qwen 4 today below.
Our picks:
- 24GB GPU and a coding or repo agent: Qwen 3.8 27B.
- 24GB GPU and you need Apache 2.0 across the whole family: Gemma 4 31B.
- 128GB-class machine: Flash Next, the closest thing to Qwen 4 you can run. Read its licence first. Ollama also lists a 105GB MLX build.
Whichever you pick, set things up so the model is a setting, not a rebuild. The day open Qwen 4 weights land, you'll want to try them on one agent before you move the rest.
On BetterClaw every agent runs on the model you choose, with your own API keys across 28+ providers at zero markup, and BetterClaw Desktop connects the Ollama models on your machine to your cloud agents. Switch one agent to Qwen4-Max, compare a week of its runs against Qwen 3.8, then decide. Plans start from $19 a month, and your first workspace starts with a 7-day free trial. See BetterClaw plans
Flash Next is the interesting one. Since Qwen says its architecture will underpin Qwen 4, running it now shows you how the next generation behaves in your own agent loops.

Running Qwen 3.8 27B in Ollama: the settings that matter
If you go with Qwen 3.8 27B, a few minutes of setup decides whether it behaves like an agent model or a slow chatbot. These are the official qwen3.8 tags in the Ollama library on 19 November 2026, all with a 256K context window and text plus image input:
| Tag | Download | Fits |
|---|---|---|
qwen3.8:27b (same as latest), 27b-q4_K_M, 27b-mtp-q4_K_M | 18GB | 24GB GPU, 24GB+ Mac |
qwen3.8:27b-mlx, 27b-nvfp4 | 18GB | Apple Silicon with 24GB+ |
qwen3.8:27b-dflash2-nvfp4 (MLX, added two weeks ago) | 19GB | Apple Silicon with 24GB+ |
qwen3.8:27b-q8_0, 27b-mtp-q8_0, 27b-mxfp8 (updated last week) | 30GB | 32GB+ GPU or Mac |
qwen3.8:27b-bf16, 27b-mlx-bf16 | 56GB | 64GB+ |
It still runs on a 16GB card, but Ollama spills layers to system RAM and generation slows sharply. Run ollama ps while the model is loaded: the PROCESSOR column shows the GPU and CPU split.
Raise the context window. Ollama does not give a model its full context by default. It allocates 4K below 24GiB of VRAM, 32K from 24 to 48GiB, and 256K only from 48GiB up, and its docs recommend at least 64K for agents and coding tools. A 4K window is the most common reason a local agent forgets its instructions or calls the wrong tool. Set it in a Modelfile:
FROM qwen3.8:27b
PARAMETER num_ctx 65536
PARAMETER num_predict 8192
PARAMETER temperature 1.0
PARAMETER top_p 0.95
PARAMETER top_k 20
Build it with ollama create qwen38-agent -f qwen38-agent.modelfile. The sampling values are the model card's settings for thinking mode. Thinking tokens count toward num_predict, so set it too low and answers stop mid-sentence. Leave the repetition penalty at 1.0, as the card recommends, rather than copying repeat_penalty 1.1 from older Qwen guides.
Switch thinking off for routine turns. Thinking is on by default. For classification, extraction and routing it mostly costs time and tokens. Use --think=false on the command line, /set nothink in a session, or "think": false on /api/chat. With thinking off, the card recommends temperature 0.7, top_p 0.8 and presence_penalty 1.5.
Prefer the native API path. Ollama's library page now offers ollama launch openclaw --model qwen3.8 and ollama launch hermes --model qwen3.8 to wire up an agent in one command. If you connect a framework by hand, use its native Ollama provider on /api/chat rather than the OpenAI-compatible /v1 endpoint where you have the choice: an early Ollama report of /v1/chat/completions hanging on qwen3.8:27b, while /api/chat worked, was closed without a fix.
If you're still on Qwen 3.7. There is nothing to pull. Qwen3.7-Max and Qwen3.7-Plus were hosted-only, ollama.com/library/qwen3.7 still returns a 404, and the open-weight line skipped straight to 3.8. If you used 3.7 through an API, Qwen 3.8 27B is the local replacement, and qwen3.6:27b remains in the Ollama library as a fallback.
Run Qwen 4 on Ollama
Status 19 November 2026: not possible yet. There is no qwen4 page in the Ollama library, because there are no Qwen 4 weights to package. Two things people type that are not Qwen 4: qwen:4b and qwen3.5:4b are 4-billion-parameter models from older generations; the "4" is a size, not a version.
What you can run today that is closest to Qwen 4:
| Want | Pull | Size | Why |
|---|---|---|---|
| The Qwen 4 architecture, locally | ollama run qwen3.8-flash-next | 105GB to 120GB, needs a 128GB-class machine | Qwen calls Flash-Next "an experimental preview of the architecture that will underpin Qwen4"; llama.cpp and vLLM name it qwen4_exp |
| The best Qwen on a 24GB card | ollama run qwen3.8:27b | 18GB | Apache 2.0, tools, vision and thinking, 256K context; settings above |
| Qwen 4 itself, through an API, from a local agent | OpenRouter or Alibaba Cloud key, model id qwen4-max | — | See Try Qwen 4 today below |
Why day-one support is likely. Ollama has to support an architecture before it can list a model. Qwen3.8-Flash-Next already runs in Ollama, and its architecture is in llama.cpp as qwen4_exp, with performance work landing this week in vLLM and SGLang (llama.cpp also has an open "Get ready for qwen4" issue). If the open Qwen 4 sizes share that design, most of the engine work is done. That is our reading of the repositories, not a promise from Qwen or Ollama.
Release-day check (we run this every weekday; copy it):
# 1. Is there an official library page yet? (200 means yes)
curl -s -o /dev/null -w "%{http_code}\n" https://ollama.com/library/qwen4
# 2. Pull the exact tag shown on that page, then look for "tools" under Capabilities
ollama pull qwen4
ollama show qwen4
The qwen4 tag is a guess until the library page exists. Copy the real tag from the page, then check ollama show lists tools under Capabilities before you connect an agent. Our list of Ollama models with tool calling support explains what to look for.
Try Qwen 4 today: API only, and what it costs
If you want Qwen 4 in an agent this week, it is an API call to qwen4-max. Three routes, all checked 19 November 2026:
| Route | Model id | Price per 1M in / out | Notes |
|---|---|---|---|
| Alibaba Cloud Model Studio (Singapore, Virginia, Frankfurt, Tokyo) | qwen4-max | $1.50 / $4.50; cache hit $0.075 | 1M free tokens for 90 days; OpenAI- and Anthropic-compatible endpoints |
| OpenRouter | qwen/qwen4-max | $1.425 / $4.275 | One key; tools and reasoning_effort supported |
| DeepInfra, Vercel AI Gateway, Novita, Venice, Kilo | qwen4-max variants | about $1.41–1.60 / $4.24–4.80 | Listed on day one |
Four settings that bite agent builders, straight from Alibaba's model page:
- Thinking is on by default at
reasoning_effort: high. For routing, classification and extraction, turn thinking off; you pay for every thinking token. max_tokensnow caps thinking and answer together (262,144 in thinking mode). Set it high enough or answers stop mid-sentence: the same failure mode asnum_predicton Qwen 3.8 locally.preserve_thinkingis on, so reasoning from earlier turns is carried forward. In a 30-step loop that is context you pay for. Our Qwen 3.8 27B tool-calling fix covers the same setting.- It is slow per turn. Artificial Analysis measured 146 seconds to first answer token and about 31 minutes per Index task. Fine for an overnight job, painful for a chat agent. Compare cost per task, not per token.
On BetterClaw, pick OpenRouter or Model Studio (Alibaba) as the provider on one agent's Model tab, paste your key and choose qwen4-max (if it isn't listed yet, OpenAI Custom with your provider's endpoint and that model id works), then run it next to a Qwen 3.8 agent for a week before moving anything else. Your bill stays with the provider; BetterClaw does not mark up inference. See plans · how BYOK works · local models with BetterClaw Desktop.
What Qwen 4 means for OpenClaw and Hermes users
If your agent runs on OpenClaw, Hermes or Qwen Code, all three can already call qwen4-max. Qwen Code merged first-class support on 17 November (effort ladder, cache TTL, image input). OpenClaw and Hermes take it through any OpenAI-compatible provider; Nous Research's portal is offering Qwen4-Max free for five days at launch, according to Qwen's own announcement. Nothing changes for your local setup: there is still no Qwen 4 you can serve from Ollama, so a hybrid (Qwen4-Max for the planning step, qwen3.8:27b for the cheap tool loops) is the honest configuration until open weights land. See our OpenClaw + Ollama guide for the local half.
The model you can run beats the model you're waiting for
Qwen has now put four things in writing: Max is out, it is closed, smaller open models are coming, and there is no date.
Everything else about the open sizes, the date, the names, the licence, is still a rumour with a nice font.
So run Qwen 3.8 now, keep your agent's model a one-line change, and keep a small test set ready. Write down the five jobs your agent does most, with the answer you expect for each. That's your test set, and it takes ten minutes.
When open Qwen 4 weights ship, the team with tests ready decides in an afternoon. Everyone else spends the week reading launch posts.
Changelog
- 19 Nov 2026: Rewrote the status block: Qwen4-Max is out (API only); open weights and Ollama still pending. Added the benchmarks table, the "Run Qwen 4 on Ollama" and "Try Qwen 4 today" sections, and Ollama's new
qwen3.8:27b-dflash2-nvfp4tag. - 16 Nov 2026: Qwen4-Max released (not reflected on this page until 19 November; we missed it by three days).
- 12 Nov 2026: Added the Ollama tag table for qwen3.8.
- 11 Nov 2026: Published.
Frequently Asked Questions
Is Qwen 4 available?
Partly. Qwen4-Max has been available since 16 November 2026 as a hosted API model on QwenCloud, Alibaba Cloud Model Studio and OpenRouter. No Qwen 4 weights have been published, so you cannot download or self-host any Qwen 4 model yet. Checked 19 November 2026.
When is Qwen 4 coming out?
The flagship is out: Qwen4-Max launched 16 November 2026. For the open-weight sizes Qwen has said only that smaller models are "coming to open source", with no date. Every specific date for a Qwen 4 27B comes from forums, not Qwen.
Is there a Qwen 4 27B?
Not as of 19 November 2026. Qwen has not named a 27B; the Qwen4-27B repositories on Hugging Face are third-party placeholders with no weights. A 27B is plausible because it was the one open size in Qwen 3.5, 3.6 and 3.8, but that is a pattern, not an announcement.
Will Qwen 4 be open source?
The flagship is not: Qwen4-Max is API-only. Qwen's blog says smaller Qwen 4 models "are coming to open source, built on the same lineage". Licence per size is unknown; Qwen 3.8 27B is Apache 2.0 while Flash-Next uses the Qwen Community licence.
Qwen 4 vs Qwen 3.8, what is new?
A new MoE architecture (Gated DeltaNet plus Qwen Sparse Attention, gated residuals, n-gram embedding), 71B active parameters instead of 95B, 1M context, native video input, and 25% lower prices. On Alibaba's chart Qwen4-Max beats Qwen3.8-Max on every benchmark shown; independent indexes agree (AA 43 vs 34).
Can I run Qwen 4 in Ollama?
No. ollama.com/library/qwen4 does not exist because no Qwen 4 weights exist. Run qwen3.8:27b (18GB) or, on a 128GB machine, qwen3.8-flash-next, which previews the Qwen 4 architecture. qwen:4b is a 4-billion-parameter model, not Qwen 4.
How much does Qwen 4 cost?
Qwen4-Max is $1.50 per million input tokens and $4.50 per million output tokens on Alibaba Cloud's international endpoints, $0.075 for cached input, and about 5% less on OpenRouter. Artificial Analysis measured $3.27 per Intelligence Index task against $10.02 for Qwen3.8-Max. Checked 19 November 2026.
What should I run until Qwen 4 open weights are released?
Qwen 3.8 27B (qwen3.8:27b, 18GB) on a 24GB GPU for coding and agent work; Gemma 4 31B if you need Apache 2.0 across the whole family; Qwen3.8-Flash-Next on a 128GB-class machine. Keep the model a one-line setting so you can switch the day weights drop.




